Emulating CMAQ using deep learning: A comparative study on simulating surface NO₂, O₃, and PM_2.5 over the CONUS using...

Area of Science:

• Environmental Science
• Computer Science
• Atmospheric Chemistry

Background:

• Air quality modeling is crucial for public health and environmental monitoring.
• Traditional chemical transport models (CTMs) are computationally intensive, limiting their application for rapid assessments.
• Developing efficient emulators is essential for faster air quality predictions.

Purpose of the Study:

• To develop and validate 2D convolutional neural network (CNN)-based emulators for predicting daily mean surface concentrations of NO2, O3, and PM2.5.
• To assess the performance of these emulators against established CTMs (CMAQ) using real-world data.
• To evaluate the computational efficiency of the deep learning approach compared to traditional modeling.

Main Methods:

• Utilized a U-Net architecture for 2D CNN emulators.
• Input data included meteorology, emissions, and land surface data from the EPA's EQUATES dataset.
• CMAQ model outputs served as target data for training and validation.
• Evaluated emulator performance using indices of agreement (IOA) and spatiotemporal analysis.

Main Results:

• Emulators achieved high agreement with CMAQ simulations (IOA up to 0.95 for NO2, 0.88 for O3, 0.85 for PM2.5).
• Demonstrated consistent spatiotemporal accuracy across diverse conditions and emission patterns.
• Achieved significant speedup (1064x faster for NO2 simulation) compared to CMAQ on a CPU-GPU setup.

Conclusions:

• Deep learning emulators offer a computationally efficient alternative for air quality assessment with comparable accuracy to CTMs.
• The 2D approach shows limitations for complex species like O3 and PM2.5, suggesting potential benefits of 3D emulators.
• This technology enables faster and more accessible air quality predictions for policy and research.